EEND-vector clustering

The EEND-vector clustering (End-to-End-Neural-Diarization-vector clustering) is a speaker diarization framework that integrates two complementary major diarization approaches, i.e., traditional clustering-based and emerging end-to-end neural network-based approaches, to make the best of both worlds. In [1] it is shown that the EEND-vector clustering outperforms EEND when the recording is long (e.g., more than 5 min), while in [2] it is shown based on CALLHOME data that it outperforms x-vector clustering and EEND-EDA especially when the number of speakers in recordings is large.

This repository contains an example implementation of the EEND-vector clustering based on Pytorch to reproduce the results in [2], i.e., the CALLHOME experiments. For the trainer, we use Padertorch. This repository is implemented based on EEND and relies on some useful functions provided therein.


[1] Keisuke Kinoshita, Marc Delcroix, and Naohiro Tawara, “Integrating end-to-end neural and clustering-based diarization: Getting the best of both worlds,” Proc. ICASSP, pp. 7198–7202, 2021

[2] Keisuke Kinoshita, Marc Delcroix, and Naohiro Tawara, “Advances in integration of end-to-end neural and clustering-based diarization for real conversational speech,” Proc. Interspeech, 2021 (to appear)


 author = {Keisuke Kinoshita and Marc Delcroix and Naohiro Tawara},
 title = {Integrating End-to-End Neural and Clustering-Based Diarization: Getting the Best of Both Worlds},
 booktitle = {{ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}},
 year = {2021}

Install tools


  • CUDA Toolkit (version == 9.2, 10.1 or 10.2)

Install kaldi and python environment

cd tools
  • This command builds kaldi at tools/kaldi
    • if you want to use pre-build kaldi
      <div class="highlight highlight-source-shell position-relative" data-snippet-clipboard-copy-content="cd tools
      make KALDI=

      cd tools
      make KALDI=<existing_kaldi_root>